ESTRO 2024 - Abstract Book

S3065

Physics - Autosegmentation

ESTRO 2024

approach. The differences in SDSC between predicted ITV’s with the all-breathing phases approach and the peak-to peak approach were not statistically significant. A total of 17 volumes were incorrectly identified as false positives: the algorithm contoured them as GTV, but analysis by a radiation oncologist revealed that these volumes represented fluid in the lungs (3 cases), aorta (3), vessels (7 cases) or other lung lesions (4).

Mean Surface dice 3mm (SD)

Median

Hausdorff

Struct

Algorithm

Mean Dice (SD)

95th (SD)

GTV 50%

Swin UNetR

0.70+-0.16

0.76+- 0.12

3.79+-2.60

GTV 50%

DynUNet

0.68+-0.17

0.74+-0.11

3.48+-2.15

GTV 50%

Swin+Dyn

0.76+-0.13

0.80+-0.11

3.45+-2.29

ITV (10BP)

Swin+Dyn

0.65+-0.09

0.83+-0.08

5.57+-2.17

ITV (2BP)

Swin+Dyn

0.69+-0.09

0.85+-0.07

5.04+-2.22

Table 1: Table with test set metrics. The first section describes the metrics of the performance of the GTV segmentation in the 50% BP. The second section describes the ITV volumes by all breathing phases (all BP) and peak to-peak (peaks).

Conclusion:

The deep learning algorithm (SwinUNETR + DynUNET) was able to delineate a precise and robust ITV, by assembling the automatic delineated GTV from all breathing phases on a 4D-CT scan. The findings could potentially be used to speed-up and reduce variability in the tumor contouring process for early-stage lung cancer patients undergoing SBRT.

Keywords: 4DCT Imaging, Deep Learning, Tumor Contouring

References:

[1] Ezhil, M., Vedam, S., Balter, P. et al. Determination of patient-specific internal gross tumor volumes for lung cancer using four-dimensional computed tomography. Radiat Oncol 4, 4 (2009). https://doi.org/10.1186/1748-717X-4-4

1574

Digital Poster

Deep learning nnU-Net for auto-contouring in cervical brachytherapy, an African country’s experience

Didier Raphael Roger Duprez 1,2 , Christoph Trauernicht 1,2